Replace built-in edit tool with custom hashline edit tool

[IshanArya]

Problem

The hashline plugin annotates Read tool output with #HL line refs (e.g. 12#A3F#9BC) and REV tokens so the LLM can make precise, hash-verified edits. Previously, we tried to extend the native Edit tool's JSON Schema at runtime via the tool.definition hook, adding operations/fileRev/safeReapply fields alongside the existing oldString/newString.

This approach failed. Session export inspection confirmed that every edit call used oldString/newString -- the schema extension either didn't propagate to the LLM or wasn't persuasive enough to override trained behavior. The hashline refs were generated but never consumed.

Solution

Replace the built-in edit tool entirely with a custom tool (.opencode/tools/edit.ts) that only accepts hashline-style args. Per the OpenCode docs: "If a custom tool uses the same name as a built-in tool, the custom tool takes precedence."

The custom tool schema has no oldString/newString. The LLM must use operations[] with hash-anchored refs from read output. This matches the approach used by oh-my-openagent, which successfully steers LLMs toward hash-anchored edits by removing the native fallback path.

Changes

Created

• .opencode/tools/edit.ts -- Custom edit tool with hashline-native schema (filePath, fileRev?, operations[]). Calls runHashlineOperationsDetailed with dryRun: false to apply edits directly. Includes a concise tool description with workflow, rules, operation types, examples, and error recovery.

Modified

• .opencode/plugins/hashline-hooks.ts -- Removed ~190 lines of dead code:
  • Removed the tool.definition hook's edit schema extension block
  • Removed translateHashlineEditArgs and all supporting functions (hasHashlineEditShape, toHashlineOperations, firstString, firstBoolean, isNativeEditTool)
  • Simplified tool.execute.before to only strip hashline prefixes from content fields (write, patch, apply_patch still need this)
  • Cleaned up unused imports (mapOperationInput, runHashlineOperationsDetailed, HashlineOperationInput)
• .opencode/plugins/hashline-routing.ts -- Removed "edit" from the known-tools set and removed all edit-specific snake_case-to-camelCase arg normalization (the custom tool defines its own schema)
• test/hashline-hardening.test.mjs -- Updated tests to verify the hook no longer modifies the edit tool's schema or description

Removed

• .opencode/tools/resolve-hash-edit.ts -- Deleted (backup in tools_disabled/). This MCP tool was a workaround that translated hashline operations to oldString/newString in a separate tool call. Redundant now that the edit tool handles hashline operations directly.

What stays the same

• Read tool -- Still uses the tool.execute.after hook to annotate output with #HL refs. Not replaced.
• Write/patch/apply_patch tools -- Still native, still get hashline prefix stripping via tool.execute.before
• System prompt injection -- experimental.chat.system.transform hook still injects hashline workflow guidance
• hashline-core.ts -- Core engine unchanged
• Ref format -- Still LINE#HASH#ANCHOR (3-4 hex chars), more collision-resistant than alternatives

How to verify

See specs/verify-hashline-edit-schema.md for the full test procedure. Quick version:

1. Start a new OpenCode session
2. Ask the LLM to read a file, then edit it
3. Export the session and inspect edit tool calls:

opencode export <sessionID> > session.json
node -e "
const d = require('./session.json');
for (const m of d.messages) {
  for (const p of (m.parts || [])) {
    if (p.type === 'tool' && p.tool === 'edit') {
      const input = p.state?.input || {};
      const style = 'operations' in input ? 'HASHLINE' : 'NATIVE';
      console.log(style, Object.keys(input));
    }
  }
}
"

All edit calls should print HASHLINE.

reference: #3